【开源实习】blip模型微调 (#1896)

Author: Alemax067

llm/finetune/blip/README.md

@@ -0,0 +1,49 @@
+# FineTune BLIP
+- reference [repo](https://github.com/eeshashetty/captionary-api)
+
+## Requirements
+- python 3.9
+- mindspore 2.3.1
+- mindnlp 0.4.1
+
+## args for training the model
+- args.device_target : Ascend
+- args.device_id 
+- args.model_name_or_path : 'Salesforce/blip-image-captioning-base' or the path to the model
+- args.dataset_name_or_path : 'eeshclusive/captionary-dataset' or the path to the data directory
+- args.batch_size : batch size
+- args.max_eps : maximum number of epochs
+- args.save_path : path to save the model, if not provided the model will not be saved, such as './outputs/'
+
+## Results
+### my results on mindspore
+20 epochs:  
+- train loss: 0.0132  
+- val loss: 0.0126
+
+requirements:
+- Ascend 910B
+- Python 3.9
+- MindSpore 2.3.1
+- MindNLP 0.4.1
+
+### my results on pytorch
+10 epochs:  
+- train loss: 0.0135  
+- val loss: 0.0125
+
+requirements:
+- GPU 1080ti
+- CUDA 11.1.1
+- Python 3.9
+- Pytorch 1.10.2
+- Transformers 4.45.2
+
+### Original results from the repo
+20 epochs:  
+- train loss: 1.3579
+- val loss: 1.3584
+
+### 其他
+- 训练loss可视化见results_visible.ipynb文件   
+- 愿仓库的损失不知为何特别高，复现时训练参数保持一致，但pytorch开启了混合精度，而mindnlp暂不支持，所以pytorch训练收敛的更快一些